TOPLINE:

A new electronic health record (EHR)-based model called PRIME identified individuals who had a sevenfold higher risk of developing pancreatic cancer, vs the population norm. Validated across more than 11 million individuals, the model achieved strong discrimination, with an area under the curve (AUC) of 0.75.

METHODOLOGY:

• Pancreatic ductal adenocarcinoma (PDAC) is the third leading cause of cancer-related mortality in the US, with most cases of the disease arising sporadically. Early detection remains a critical challenge, as existing blood biomarkers are not specific enough. A multistage approach to screening, beginning with EHR data then moving to biomarker testing in higher-risk patients, may offer an alternative.
• This study developed and externally validated PRIME, an EHR-based model incorporating routinely available PDAC risk factors. The model was trained using de-identified EHRs from nearly 5 million patients aged 40 years and older from 23 US healthcare systems. Validation was conducted using data from 5.6 million patients at 31 additional US health systems, and from nearly 500,000 UK Biobank participants.
• The final model incorporated 19 predictors, including history of pancreatitis, gastrointestinal disorders, prior cancers, type 2 diabetes, elevated aspartate aminotransferase levels, smoking, non-type-O blood, and male sex.
• Model performance was assessed using time-dependent AUC and calibration metrics at 24 and 36 months, with a mean follow-up of 5.4 years in the training cohort and 3.9 years in the validation cohort.

TAKEAWAY:

• In the US validation cohort, PRIME achieved a 36-month time-dependent AUC of 0.75 (95% CI, 0.75-0.75). Calibration indicated close agreement between observed and predicted risks (36-month slope, 1.04).
• People in the top 1% of the PRIME-predicted risk had at least a 0.61% predicted risk of developing PDAC over 36 months. Compared with the population average, those individuals were over 7 times more likely to develop the disease (hazard ratio [HR] 7.63) — on par with people who carry pathogenic germline variants, the authors noted. Individuals in the 95th to 99th percentile had a more than fourfold increase in risk (HR, 4.40).
• At 36 months, the positive predictive value (PPV) in the top 1% of predicted risk was 0.78% (1 case per 128 patients). In the top 10% of predicted risk, PPV was 0.42% (1 case per 236 patients). Overall, people in the top 1% and 10% of predicted risk accounted for 5.8% and 31.5% of all PDAC cases at 36 months.
• In the UK Biobank validation cohort, PRIME achieved a 36-month AUC of 0.71 (95% CI, 0.71-0.71) without recalibration, with modest overprediction of risk (36-month slope, 0.87).

IN PRACTICE:

" In a multistage case-finding framework, individuals with PRIME risk estimates exceeding a predefined absolute risk threshold would be candidates for downstream biomarker testing," the study authors wrote. "Patients with positive biomarker test results would then undergo imaging, similar to PDAC surveillance strategies for [pathogenic germline variant] carriers."

Future studies, they concluded, "should assess the impact of this multistage approach on downstream testing burden, stage shift, resectability, and survival." 

SOURCE:

The study was led by Lucas A. Mavromatis, ScB, and Morgan E. Grams, MD, PhD, both of New York University Grossman School of Medicine in New York City. It was published online in JAMA Oncology.

LIMITATIONS :

Some PDAC risk factors, including family history and timing of type 2 diabetes onset, are poorly captured in EHRs and were not included in the model. The use of race and ethnicity in clinical algorithms remains controversial, but their removal led to underprediction of risk among Black individuals in this study. Despite good model performance, most PDAC cases occurred outside the highest predicted risk strata, underscoring the need for more specific noninvasive biomarkers.

DISCLOSURES:

This study was funded by the National Institute of Diabetes and Digestive and Kidney Diseases. Grams disclosed support from Vertex outside the submitted work, and a co-author disclosed license and royalty interests from Exact Sciences during the conduct of the study.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.